Diagnostic performance of multimodal ultrasound-based deep learning models in differentiating benign and malignant...

Multimodal ultrasound (US), including superb microvascular imaging (SMI) and shear-wave elastography (SWE), provides complementary information on tumor angiogenesis and tissue stiffness, which may enhance the diagnostic performance of deep learning (DL) models for thyroid nodules. However, current evidence remains limited and lacks comparative evaluation across different DL architectures. This study aimed to explore the performance of different DL models for differentiating benign and malignant thyroid nodules based on multimodal US images.

This retrospective study involved 735 patients with surgically or pathologically confirmed thyroid nodules. A total of 15,373 multimodal US images, including B-mode (both longitudinal and transverse views), SMI, and SWE, were randomly divided at the image level into training (N = 11,530) and validation (N = 3,843) cohorts at a 3:1 ratio. Four convolutional neural networks, including ResNet50, DenseNet121, VGG16, and GoogLeNet, were trained and validated. The diagnostic performance of these models was compared with junior, intermediate, and senior radiologists. Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize the model's areas of focus.

In the validation cohort, the ResNet50 model achieved the highest diagnostic performance [area under the curve (AUC): 0.931], followed by DenseNet121 (AUC = 0.857), VGG16 (AUC = 0.846), and GoogLeNet (AUC = 0.811). Delong's test showed that the AUC of ResNet50 model was higher than that of the other models (all P < 0.001). Based on calibration analysis and the Hosmer-Lemeshow test, the overall calibration performance of the models was considered acceptable. The decision curve analysis suggested that ResNet50 provided the highest net clinical benefit. The diagnostic performance of the ResNet50 model (accuracy: 0.871) was numerically better than junior radiologists (accuracy: 0.810), comparable to intermediate radiologists (accuracy: 0.886), and lower than senior radiologists (accuracy: 0.946). Grad-CAM visualization suggested that the ResNet50 model mainly focused on clinically relevant thyroid nodule regions, with some misclassified cases showing excessive attention to local features.

The multimodal US-based DL models achieve satisfactory performance in differentiating benign and malignant thyroid nodules, with the ResNet50 model possessing the highest performance, which may be comparable to intermediate radiologists.